A prior for shape-from-contour

Purpose. Shape-from-contour is ambiguous. In particular, a given line-drawing can depict either a concave or a convex surface. We investigate the biases that human observers have for one or the other interpretation. Methods. The stimuli were images of smoothly folded patches, painted with parallel stripes and illuminated from the side. Two levels of contour curvature and two levels of ambient illumination were used. Each image was then presented at a random orientation in the image plane. Subjects were asked whether the fold appeared concave or convex. Results. Shading had little or no effect on the results. On the other hand, subjects were far more likely to perceive the stimulus as convex if the curved contours were also convex in the image. Conclusions. There is an oft-cited observer bias that the light source comes from above in interpreting shading patterns. That bias had no observable correlate in our results. On the other hand, our data can be analyzed as resulting from an observer bias reflecting a prior used during the estimation of shape from contour. Our observers appear to assume that the surface is tilted so that the surface normal is pointed upward. This prior is loosely consistent with the intuitive idea that objects are most often seen from "above".